MiniMax H3: 33-billion-parameter 2K video model with audio runs locally on a 12 GB graphics card via ComfyUI
MiniMax's H3 video model, whose weights were opened on 3 August, is being run on consumer graphics cards without cloud or API. According to the ComfyUI team, the memory footprint fell from 123.6 GB to 42.5 GB and the model runs on a card like the RTX 3060. The license is valid worldwide, except in t
A month after its weights opened on Hugging Face on 3 August, MiniMax's video model H3, introduced on 31 July, runs locally on consumer graphics cards. The 33-billion-parameter model generates 2K video with audio, up to 15 seconds, without cloud or API, on a 12 GB gaming card like the RTX 3060, ComfyUI's example; ComfyUI added support the day the weights opened. Social media users report the same.
What the model does
According to MiniMax, H3 is a “general-purpose, multimodal generation model” that understands text, image, video and audio in one context and generates video with native stereo audio. Two checkpoints: FL2VA (text and first/last frame) and Ref2VA (reference image, video or audio). Output: 4 to 15 seconds, 24 fps, 768-pixel short side by default, up to 2K.
How memory shrank
The memory figures are ComfyUI's, not MiniMax's: the footprint fell from 123.6 GB at full precision to 42.5 GB in the smallest variants, down 66 percent. Modulation weights, roughly 40 percent of parameters, were pruned and replaced by a functionally equivalent lookup table; weights ship with int8 “convrot” quantization; custom kernels cut peak VRAM at inference. With dynamic VRAM offloading the model runs, in ComfyUI's words, “on a GPU like the RTX 3060”. ComfyUI 0.30.0 or later is needed.
Who the license covers
The “MiniMax H3 Community License Agreement” excludes the EU, UK, Republic of Korea and US; those inside can apply at api@minimax.io. Turkey, Arab countries and China are not excluded. Commercial use is allowed if the product interface visibly displays “MiniMax H3”; products earning over 20 million dollars a year need separate written permission. Prohibited uses include military purposes and election misinformation; machine-generated content must be disclosed publicly, and outputs cannot train competing models.
For the corporate reader
For organizations keeping video generation, and their data, in-house, this is the first realistic option on consumer hardware. But a 42.5 GB footprint is not “runs on a 12 GB card”: offloading trades speed for memory; measure on your own hardware first. Those operating in the US or EU must apply for a license.
Sources
- MiniMax announcement (31 July): https://www.minimax.io/blog/minimax-h3
- Hugging Face model card, LICENSE: https://huggingface.co/MiniMaxAI/MiniMax-H3
- GitHub: https://github.com/MiniMax-AI/MiniMax-H3
- ComfyUI blog (3 August): https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui
- ComfyUI docs: https://docs.comfy.org/tutorials/video/minimax/minimax-h3
